On June 9, 2026, Anthropic released Claude Fable 5 — a model that instantly reset the bar for publicly available AI. Alongside it came Claude Mythos 5, a sibling model with even fewer restrictions, reserved for a small set of vetted institutions. The two share the exact same underlying architecture. The difference? Fable 5 ships with a safety classifier designed for public deployment; Mythos 5 does not.
This was the first time Anthropic opened a Mythos-class model to the general public. And for four days, it was the strongest AI any ordinary user could access. Then the US government stepped in.
What Makes Fable 5 Different?
Fable 5 is not a small iteration. It arrived just 11 days after Claude Opus 4.8, yet the performance gap is closer to a generational leap than a point release. Anthropic describes it as excelling at "longer, more complex tasks" — and the benchmark data backs that up dramatically.
The model can process millions of tokens in a single context window and uses an internal note-taking mechanism to track its own reasoning across extended sessions. This means it doesn't lose the plot halfway through a complex project. It remembers. It plans. It executes.
Benchmark Performance: By the Numbers
The numbers are striking. On SWE-Bench Pro, the hardest real-world software engineering benchmark available, here's how the landscape looked on June 10:
| Model | SWE-Bench Pro Score |
|---|---|
| Claude Fable 5 | 80.3% |
| Claude Opus 4.8 | 69.2% |
| Claude Mythos Preview | 77.8% |
| GPT-5.5 | 58.6% |
| Gemini 3.1 Pro | 54.2% |
An 11-percentage-point lead over the next-best public model — and it even beat Anthropic's own restricted Mythos Preview (77.8%). That kind of single-generation jump is almost unheard of.
On FrontierCode Diamond, a benchmark for elite competitive programming problems, the gap was even wider:
| Model | FrontierCode Diamond |
|---|---|
| Claude Fable 5 | 29.3% |
| Claude Opus 4.8 | 13.4% |
| GPT-5.5 | 5.7% |
Fable 5 more than doubled Opus 4.8's score and outperformed GPT-5.5 by over 5x. These are problems that stump most human engineers.
Real-World Capabilities
Benchmarks tell one story. What Fable 5 did in the wild tells another.
Stripe's 50-Million-Line Migration
Payment giant Stripe ran an early test: migrate a 50-million-line Ruby codebase. Fable 5 completed the entire migration in one day. The same project had been estimated to take a human engineering team two months. That's roughly a 60x speedup on a real production workload.
Pokémon FireRed — Screenshots Only
Fable 5 played through Pokémon FireRed using nothing but raw screenshots as input. No API, no game-specific tools, no extra instrumentation. It read the pixels, figured out the game mechanics, and played to completion. This was a long-standing informal stress test for vision-capable AI agents — and Fable 5 was the first to clear it purely visually.
Ethan Mollick's 9-Hour Autonomous Research Project
Wharton professor Ethan Mollick fed Fable 5 a 15-page requirements document. Nine hours later, he had a finished deliverable. The model autonomously spawned sub-agents, conducted research, wrote drafts, edited, scrapped bad approaches, and started over — all with zero human intervention. This is the kind of long-horizon autonomous work that previous models couldn't sustain.
Safety Architecture: The Automatic Downgrade
Fable 5's safety system is unusual. Rather than simply refusing risky queries, it employs an automatic safety downgrade mechanism. When the model's safety classifier detects a request involving high-risk domains — cybersecurity, biochemical topics, model distillation — it doesn't block the user. Instead, it silently switches to Claude Opus 4.8 to generate the response.
The user may never know the model changed. Anthropic reports that this classifier fires in fewer than 5% of sessions, and over 1,000 hours of external red-teaming found no universal jailbreak method.
This architecture is both clever and controversial. Clever because it preserves the user experience for legitimate queries while maintaining safety boundaries. Controversial because a silent downgrade means users can't always trust which model they're interacting with — and the classifier itself becomes a single point of policy enforcement.
Pricing and Availability (Before the Ban)
Fable 5 launched with pricing set at:
- $10 per million input tokens
- $50 per million output tokens
This is exactly double Opus 4.8's pricing ($5/$25), but less than half of Mythos Preview. For Pro, Max, Team, and Enterprise subscribers, Fable 5 was free from June 9 through June 22. After that, usage would require credits.
The model was available through:
- Claude web and mobile apps
- Claude API (
claude-fable-5) - Amazon Bedrock
- GitHub Copilot
- Microsoft Foundry
Then, on June 12 at 5:21 PM ET, everything changed.